How RL Creates General Problem Solving | Reflection AI

Reflection AI claims its 501B-parameter Beam model, built via massive RL scaling with zero distillation, shows generalizable agentic reasoning that can rival Chinese frontier open models.

• Beam specs: Reflection AI's new open-weight model has 501B total parameters, 23B active (mixture-of-experts), trained with zero distillation from other models.
• Record RL scale: Training used 10,500 GPUs for four weeks, over 100 million rollouts, and more than a billion sandbox environments—the largest documented open RL run.

Themes, sentiment and reasons are AI-written paraphrases, checked so they never quote an episode. They may contain errors. Not investment advice. Episodes remain the property of their creators.